在 ai.google.dev 上查看 | 在 Google Colab 中运行 | 在 GitHub 上查看源代码 |
在此笔记本中,您将了解如何通过 C网址 命令或 Python 请求 API 来调用 Gemini API 来开始使用 Gemini API 调整服务。在这里,您将学习如何调整 Gemini API 的文本生成服务背后的文本模型。
设置身份验证
借助 Gemini API,您可以使用自己的数据对模型进行调参。由于是您的数据和您调整的模型,因此这需要比 API 密钥提供更严格的访问权限控制。
在运行本教程之前,您需要为项目设置 OAuth。
在 Colab 中,最简单的设置方法是将 client_secret.json
文件的内容复制到 Colab 的“Secret 管理器”(位于左侧面板中的密钥图标下),并将其密钥名称为 CLIENT_SECRET
。
此 gcloud 命令会将 client_secret.json
文件转换为可用于向服务进行身份验证的凭据。
try:
from google.colab import userdata
import pathlib
pathlib.Path('client_secret.json').write_text(userdata.get('CLIENT_SECRET'))
# Use `--no-browser` in colab
!gcloud auth application-default login --no-browser --client-id-file client_secret.json --scopes='https://www.googleapis.com/auth/cloud-platform,https://www.googleapis.com/auth/generative-language.tuning'
except ImportError:
!gcloud auth application-default login --client-id-file client_secret.json --scopes='https://www.googleapis.com/auth/cloud-platform,https://www.googleapis.com/auth/generative-language.tuning'
You are authorizing client libraries without access to a web browser. Please run the following command on a machine with a web browser and copy its output back here. Make sure the installed gcloud version is 372.0.0 or newer. gcloud auth application-default login --remote-bootstrap="https://accounts.google.com/o/oauth2/auth?response_type=code&client_id=87071151422-n1a3cb6c7fvkfg4gmhdtmn5ulol2l4be.apps.googleusercontent.com&scope=https%3A%2F%2Fwww.googleapis.com%2Fauth%2Fcloud-platform+https%3A%2F%2Fwww.googleapis.com%2Fauth%2Fgenerative-language.tuning&state=QIyNibWSaTIsozjmvZEkVBo6EcoW0G&access_type=offline&code_challenge=76c1ZiGvKN8cvlYfj3BmbCwE4e7tvrlwaX3REUX25gY&code_challenge_method=S256&token_usage=remote" Enter the output of the above command: https://localhost:8085/?state=QIyNibWSaTIsozjmvZEkVBo6EcoW0G&code=4/0AeaYSHBKrY911S466QjKQIFODoOPXlO1mWyTYYdrbELIDV6Hw2DKRAyro62BugroSvIWsA&scope=https://www.googleapis.com/auth/cloud-platform%20https://www.googleapis.com/auth/generative-language.tuning Credentials saved to file: [/content/.config/application_default_credentials.json] These credentials will be used by any library that requests Application Default Credentials (ADC).
设置变量
CURL
为重复值设置变量,以用于其余 REST API 调用。该代码使用 Python os
库来设置可在所有代码单元中访问的环境变量。
这特定于 Colab 笔记本环境。下一个代码单元中的代码相当于在 bash 终端中运行以下命令。
export access_token=$(gcloud auth application-default print-access-token)
export project_id=my-project-id
export base_url=https://generativelanguage.googleapis.com
import os
access_token = !gcloud auth application-default print-access-token
access_token = '\n'.join(access_token)
os.environ['access_token'] = access_token
os.environ['project_id'] = "[Enter your project-id here]"
os.environ['base_url'] = "https://generativelanguage.googleapis.com"
Python
access_token = !gcloud auth application-default print-access-token
access_token = '\n'.join(access_token)
project = '[Enter your project-id here]'
base_url = "https://generativelanguage.googleapis.com"
导入 requests
库。
import requests
import json
列出经调参的模型
列出可用的已调参模型,验证您的身份验证设置。
CURL
curl -X GET ${base_url}/v1beta/tunedModels \
-H "Content-Type: application/json" \
-H "Authorization: Bearer ${access_token}" \
-H "x-goog-user-project: ${project_id}"
Python
headers={
'Authorization': 'Bearer ' + access_token,
'Content-Type': 'application/json',
'x-goog-user-project': project
}
result = requests.get(
url=f'{base_url}/v1beta/tunedModels',
headers = headers,
)
result.json()
创建经调参的模型
如需创建经调整的模型,您需要将数据集传递给 training_data
字段中的模型。
在此示例中,您将调整模型以生成序列中的下一个数字。例如,如果输入为 1
,则模型应输出 2
。如果输入为 one hundred
,输出应为 one hundred one
。
CURL
curl -X POST $base_url/v1beta/tunedModels \
-H 'Content-Type: application/json' \
-H "Authorization: Bearer ${access_token}" \
-H "x-goog-user-project: ${project_id}" \
-d '
{
"display_name": "number generator model",
"base_model": "models/gemini-1.0-pro-001",
"tuning_task": {
"hyperparameters": {
"batch_size": 2,
"learning_rate": 0.001,
"epoch_count":5,
},
"training_data": {
"examples": {
"examples": [
{
"text_input": "1",
"output": "2",
},{
"text_input": "3",
"output": "4",
},{
"text_input": "-3",
"output": "-2",
},{
"text_input": "twenty two",
"output": "twenty three",
},{
"text_input": "two hundred",
"output": "two hundred one",
},{
"text_input": "ninety nine",
"output": "one hundred",
},{
"text_input": "8",
"output": "9",
},{
"text_input": "-98",
"output": "-97",
},{
"text_input": "1,000",
"output": "1,001",
},{
"text_input": "10,100,000",
"output": "10,100,001",
},{
"text_input": "thirteen",
"output": "fourteen",
},{
"text_input": "eighty",
"output": "eighty one",
},{
"text_input": "one",
"output": "two",
},{
"text_input": "three",
"output": "four",
},{
"text_input": "seven",
"output": "eight",
}
]
}
}
}
}' | tee tunemodel.json
{ "name": "tunedModels/number-generator-model-dzlmi0gswwqb/operations/bvl8dymw0fhw", "metadata": { "@type": "type.googleapis.com/google.ai.generativelanguage.v1beta.CreateTunedModelMetadata", "totalSteps": 38, "tunedModel": "tunedModels/number-generator-model-dzlmi0gswwqb" } } % Total % Received % Xferd Average Speed Time Time Time Current Dload Upload Total Spent Left Speed 100 2280 0 296 100 1984 611 4098 --:--:-- --:--:-- --:--:-- 4720
Python
operation = requests.post(
url = f'{base_url}/v1beta/tunedModels',
headers=headers,
json= {
"display_name": "number generator",
"base_model": "models/gemini-1.0-pro-001",
"tuning_task": {
"hyperparameters": {
"batch_size": 4,
"learning_rate": 0.001,
"epoch_count":5,
},
"training_data": {
"examples": {
"examples": [
{
'text_input': '1',
'output': '2',
},{
'text_input': '3',
'output': '4',
},{
'text_input': '-3',
'output': '-2',
},{
'text_input': 'twenty two',
'output': 'twenty three',
},{
'text_input': 'two hundred',
'output': 'two hundred one',
},{
'text_input': 'ninety nine',
'output': 'one hundred',
},{
'text_input': '8',
'output': '9',
},{
'text_input': '-98',
'output': '-97',
},{
'text_input': '1,000',
'output': '1,001',
},{
'text_input': '10,100,000',
'output': '10,100,001',
},{
'text_input': 'thirteen',
'output': 'fourteen',
},{
'text_input': 'eighty',
'output': 'eighty one',
},{
'text_input': 'one',
'output': 'two',
},{
'text_input': 'three',
'output': 'four',
},{
'text_input': 'seven',
'output': 'eight',
}
]
}
}
}
}
)
operation
<Response [200]>
operation.json()
{'name': 'tunedModels/number-generator-wl1qr34x2py/operations/41vni3zk0a47', 'metadata': {'@type': 'type.googleapis.com/google.ai.generativelanguage.v1beta.CreateTunedModelMetadata', 'totalSteps': 19, 'tunedModel': 'tunedModels/number-generator-wl1qr34x2py'} }
使用经调整的模型的名称设置一个变量,以用于其余调用。
name=operation.json()["metadata"]["tunedModel"]
name
'tunedModels/number-generator-wl1qr34x2py'
获取经调参的模型状态
模型的状态在训练期间设置为 CREATING
,并在训练完成后更改为 ACTIVE
。
CURL
下面的一些 Python 代码用于从响应 JSON 中解析生成的模型名称。如果您在终端中运行此命令,可以尝试使用 bash JSON 解析器来解析响应。
import json
first_page = json.load(open('tunemodel.json'))
os.environ['modelname'] = first_page['metadata']['tunedModel']
print(os.environ['modelname'])
tunedModels/number-generator-model-dzlmi0gswwqb
使用模型名称执行另一个 GET
请求,以获取包含状态字段的模型元数据。
curl -X GET ${base_url}/v1beta/${modelname} \
-H 'Content-Type: application/json' \
-H "Authorization: Bearer ${access_token}" \
-H "x-goog-user-project: ${project_id}" | grep state
"state": "ACTIVE", % Total % Received % Xferd Average Speed Time Time Time Current Dload Upload Total Spent Left Speed 100 5921 0 5921 0 0 13164 0 --:--:-- --:--:-- --:--:-- 13157
Python
tuned_model = requests.get(
url = f'{base_url}/v1beta/{name}',
headers=headers,
)
tuned_model.json()
以下代码每 5 秒检查一次状态字段,直到它不再处于 CREATING
状态。
import time
import pprint
op_json = operation.json()
response = op_json.get('response')
error = op_json.get('error')
while response is None and error is None:
time.sleep(5)
operation = requests.get(
url = f'{base_url}/v1/{op_json["name"]}',
headers=headers,
)
op_json = operation.json()
response = op_json.get('response')
error = op_json.get('error')
percent = op_json['metadata'].get('completedPercent')
if percent is not None:
print(f"{percent:.2f}% - {op_json['metadata']['snapshots'][-1]}")
print()
if error is not None:
raise Exception(error)
100.00% - {'step': 19, 'epoch': 5, 'meanLoss': 1.402067, 'computeTime': '2024-03-14T15:11:23.766989274Z'}
运行推理
调优作业完成后,您可以使用它通过文本服务生成文本。
CURL
尝试输入罗马数字,例如 63 (LXIII):
curl -X POST $base_url/v1beta/$modelname:generateContent \
-H 'Content-Type: application/json' \
-H "Authorization: Bearer ${access_token}" \
-H "x-goog-user-project: ${project_id}" \
-d '{
"contents": [{
"parts": [{
"text": "LXIII"
}]
}]
}' 2> /dev/null
{ "candidates": [ { "content": { "parts": [ { "text": "LXIV" } ], "role": "model" }, "finishReason": "STOP", "index": 0, "safetyRatings": [ { "category": "HARM_CATEGORY_SEXUALLY_EXPLICIT", "probability": "NEGLIGIBLE" }, { "category": "HARM_CATEGORY_HATE_SPEECH", "probability": "NEGLIGIBLE" }, { "category": "HARM_CATEGORY_HARASSMENT", "probability": "NEGLIGIBLE" }, { "category": "HARM_CATEGORY_DANGEROUS_CONTENT", "probability": "NEGLIGIBLE" } ] } ], "promptFeedback": { "safetyRatings": [ { "category": "HARM_CATEGORY_SEXUALLY_EXPLICIT", "probability": "NEGLIGIBLE" }, { "category": "HARM_CATEGORY_HATE_SPEECH", "probability": "NEGLIGIBLE" }, { "category": "HARM_CATEGORY_HARASSMENT", "probability": "NEGLIGIBLE" }, { "category": "HARM_CATEGORY_DANGEROUS_CONTENT", "probability": "NEGLIGIBLE" } ] } }
模型的输出不一定正确。如果调整后的模型没有达到您所需的标准,您可以尝试添加更多高质量的示例、调整超参数或向示例添加前导。您甚至可以根据您创建的第一个模型再创建一个经调参的模型。
如需获得有关如何提高性能的更多指导,请参阅调参指南。
Python
尝试输入日语数字,例如 6 (6):
import time
m = requests.post(
url = f'{base_url}/v1beta/{name}:generateContent',
headers=headers,
json= {
"contents": [{
"parts": [{
"text": "六"
}]
}]
})
import pprint
pprint.pprint(m.json())
{'candidates': [{'content': {'parts': [{'text': '七'}], 'role': 'model'}, 'finishReason': 'STOP', 'index': 0, 'safetyRatings': [{'category': 'HARM_CATEGORY_SEXUALLY_EXPLICIT', 'probability': 'NEGLIGIBLE'}, {'category': 'HARM_CATEGORY_HATE_SPEECH', 'probability': 'NEGLIGIBLE'}, {'category': 'HARM_CATEGORY_HARASSMENT', 'probability': 'LOW'}, {'category': 'HARM_CATEGORY_DANGEROUS_CONTENT', 'probability': 'NEGLIGIBLE'}]}], 'promptFeedback': {'safetyRatings': [{'category': 'HARM_CATEGORY_SEXUALLY_EXPLICIT', 'probability': 'NEGLIGIBLE'}, {'category': 'HARM_CATEGORY_HATE_SPEECH', 'probability': 'NEGLIGIBLE'}, {'category': 'HARM_CATEGORY_HARASSMENT', 'probability': 'NEGLIGIBLE'}, {'category': 'HARM_CATEGORY_DANGEROUS_CONTENT', 'probability': 'NEGLIGIBLE'}]} }
模型的输出不一定正确。如果调整后的模型没有达到您所需的标准,您可以尝试添加更多高质量的示例、调整超参数或向示例添加前导。
总结
虽然训练数据不包含任何对罗马数字或日语数字的引用,但模型在微调后仍然能够很好地进行泛化。这样,您就可以根据自己的使用场景来微调模型。
后续步骤
如需了解如何借助适用于 Gemini API 的 Python SDK 使用调整服务,请参阅使用 Python 进行调参快速入门。如需了解如何使用 Gemini API 中的其他服务,请访问 Python 快速入门。